Building in the open.
One post at a time.

Five stories this week, one question: can you show me?

This week's school news was about money, data, workload, regulation and evidence. Every story landed on the same question, and a school needs its own answer.

4 min read

Four people asked for the same record this week

A parent, a government call for evidence, a governing board and a marking model all asked schools for the same record this week. None of them got it.

4 min read

Ask for the override number

A board cannot judge an AI tool by its policy. It can ask how often a teacher changed what the tool drafted, and the school should hold the answer itself.

4 min read

The diagnosis was doing the triage

The review says support should not wait for a diagnosis. That makes a school's own record of need the thing that triggers help, and it is mostly unwritten.

5 min read

A board can only challenge what reaches it.

Four stories from one week show the same gap in schools: the people responsible for an answer cannot see the numbers underneath it.

4 min read

Four duties, one term, and one person's memory

Somebody will ask you in December to describe a year you did not have time to write down. That is the shape of this autumn term.

The deadline was doing the work.

One strand of policy is removing the deadline that produced a current record. Another is raising how often somebody has to produce one on demand. They are the same unresolved disagreement.

Two inventories decide what a school can prove. Neither was ever anybody's job.

A week of new duties assumes schools hold two records: what was put in place for each child, and which AI tools are in use and what each was allowed to read. Most schools hold neither.

The regression we found in our own model

We trained a model to adapt lessons for SEND pupils and it got worse at noticing safeguarding disclosures. What we found, and what we changed.

The number that looked right.

Over three days I found five errors in my own evaluation code. None was in a model. Every one was in the code that measured the model.

Trained on Dawn.

Aevum's base models were trained on Dawn, the UK's fastest AI supercomputer, at the University of Cambridge — under UK law, on licence-clean data. What that means for schools.

AI can draft the marking. It cannot own the judgement.

DfE guidance permits AI for low-stakes marking and rules it out for judgements about a child's attainment. That line is the whole design brief for AI marking and feedback.

AI for SEND is not a smaller version of AI for everyone else.

Technology reaches pupils with SEND last, because it is built for the average child and adapted afterwards. What changes when you design for AAC, Makaton and EHCP learners first.

The same question, five times. It didn't agree with itself.

We asked our own child-safeguarding model the same disclosure prompt five times. Same weights, same settings, different answers. That's not a curiosity. It's the reason we test differently now.

The letters, the reports, the forms. The job around the job.

DfE guidance names routine parent letters as a reasonable use of AI. The harder question is which paperwork carries a child's name, and what that should change about where it is drafted.

Differentiation was always the hour nobody paid you for.

Producing four versions of the same worksheet is a production problem, not a teaching problem. Which part of differentiation AI can take, and which part has to stay with the teacher.

Accessible by default, not accessible on request.

A resource that has to be specially adapted for one child has already told that child something. How AI changes the economics of accessible learning materials, and where it still gets it wrong.

Thirty pupils, eleven languages, one letter home.

Machine translation is now good enough for most school communication. The remaining risks are not the ones people expect, and the data question is the one nobody asks.

Draft the plan, but keep the child in the room.

New DfE guidance lets schools use AI to draft SEND support plans while warning them to guard a child's data. On-device AI is the honest way to do both.

Sovereign AI for schools. Why your data has to stay in your building.

Banks and defence made sovereign AI a procurement requirement. Schools have the same problem — and none of the leverage.

AI will not give you your evenings back on its own.

Some tasks give back real time with AI. Others cost more to supervise than they ever save. Knowing which is which is most of the value, and nobody sells that part.

We generate our own images. We also tell you that we do.

Aevum's fifth tool is a multimedia pipeline, and it's also how we make our own brand imagery: with open models, on our own systems, always disclosed.

Not every business has a back office. Some are one person and a laptop.

Aevum also builds for the person running a business alone: one working environment for the legal, financial, customer and admin work a solo founder does at midnight.

Giving a voice when you cannot find the words.

A privacy-first, edge-first wellbeing wearable built for the Qwen Cloud Global AI Hackathon — and a live feelings card for children who cannot always find the words.

The work between the lessons. And the learners the system serves last.

A tutoring environment for the learning that happens between teacher, parent and pupil, built SEND-first, aware of AAC and Makaton from the very start.

Building what my family needed and never had.

Why I'm building Aevum Life — a free app for the people the market serves last, from a founder who grew up poor and has known exclusion from both sides.

Twelve questions to ask before you sign anything.

A school has no procurement department and a supplier has a sales team. Twelve questions that change the answer, including the ones we would rather you did not ask us.

Running a school is a second full-time job. We built for that one too.

The parts of a school that have nothing to do with teaching are their own mountain. Aevum's school operations tool is built for it, and the school keeps its data.

Your pupils are already using it. The lesson is how to think about it.

Detection does not work reliably and bans do not survive contact with a phone. What is left is teaching pupils to interrogate a system that is confident and sometimes wrong.

The lessons are the job. The evenings shouldn't be.

The first Aevum tool takes on the work that steals a teacher's evenings: the marking, the planning, the endless letters. Your pupils' work never leaves the building.

Will AI replace teachers? No. But it will change the argument.

The honest answer is no. The honest follow-up is that nobody is reassured by being told not to worry. What is actually at stake, and who decides it.

Not one big app. A small family that works together.

Aevum isn't one do-everything chatbot. It's five tools with five jobs, sharing one promise: your data stays in the building, and nothing you enter comes to us.

Is it worth the licence? A one-term check for SENCOs

A one-term check for SENCOs deciding whether an AI or edtech tool did what it claimed, for the pupils in front of them.